Dimensionality reduction for more stable vision parameter estimation

The problem of estimating parameters from data is considered for a class of multi-objective models of importance in computer vision. One previous approach for solving the problem is via the fundamental numerical scheme (FNS). Here, a more stable version of FNS is developed, with better convergence properties than the original version. The improvement in performance is achieved by reducing the original estimation problem to a couple of problems of lower dimension. By way of example, the new algorithm is applied to the problem of estimating the trifocal tensor relating three views of a scene. Experiments carried out with both synthetic and real images reveal the new estimator to be more stable compared to the original FNS method, and commensurate in accuracy with, but faster than, the gold standard maximum likelihood estimator.